Researchers introduce LuxSQA, a method for Spoken Question Answering (SQA) in Luxembourgish that leverages text-to-speech synthesis to generate training data without requiring large human-recorded corpora. The approach translates existing text-based QA resources into Luxembourgish and synthesizes spoken questions using multiple TTS systems, which are then paired with textual answers.
- The team trains a parameter-efficient SLAM-style architecture connecting a frozen Whisper encoder to frozen multilingual LLM backends via a learned projector and LoRA adapters.
- They compare MMS-TTS, Qwen3-TTS, and OmniVoice variants using single-source corpora of about 48k questions and a multi-source mix of approximately 230k questions.
- Evaluation on the LLAMA-LB-Test dataset with real Luxembourgish speaker conditions shows that multi-source and voice-design-based synthetic training configurations yield the strongest SQA performance.
- The results indicate that no-reference TTS quality scores do not monotonically predict downstream QA performance, suggesting synthetic speech should be evaluated as task-specific training data rather than solely for naturalness.
This work demonstrates that TTS can effectively provide task-specific training data for low-resource SQA settings, expanding the reach of speech-LLM methods beyond high-resource languages.